Grounding
Connecting language model outputs to verifiable, factual information sources to reduce hallucination and improve accuracy.
Techniques
RAG (Retrieval-Augmented Generation): Retrieve relevant documents and include them in the prompt. Tool use: Let the model query databases, search engines, or APIs for facts. Citation: Train models to cite their sources.
Why Grounding Matters
Language models generate plausible-sounding text that may not be factually correct. Grounding anchors responses in real data, making the model verifiable and trustworthy. It's essential for any application where accuracy matters.
Evaluation
Grounded responses can be verified against their sources. Attribution metrics measure whether the model's claims are supported by the provided context. This makes grounded systems more auditable than ungrounded generation.